~ Computer Graphics, Volume24, Number4, August 1990 The Accumulation Buffer: Hardware Support for High-Quality Rendering Paul Haeberli and Kurt Akeley Silicon Graphics ComputerSystems ABSTRACT 1. Introduction Traditional 3D graphics workstations include simplistic scan This paper describes a system architecture that supports realtime conversion hardware that samples points, lines, and polygons with generation of complex images, efficient generation of extremely a single infinitely small sample per pixel. As a result, the high-quality images, and a smooth trade-off between the two. renderings of these primitives show aliasing artifacts. While Based on the paradigm of integration, the architecture extends a increasing monitor pixel density has reduced the effects of state-of-the-art rendering system with an additional high-precision aliasing, motion, the result of increased workstation performance, image buffer. This additional buffer, called the Accumulation has increased them. The problem of aliasing remains significant Buffer, is used to integrate images that are rendered into the in workstation rendering architectures. framebuffer. While originally conceived as a solution to the Contemporary workstations have attempted to solve the problem problem of aliasing, the Accumulation Buffer provides a general of aliasing using a variety of architectures. Several vendors offer solution to the problems of motion blur and depth-of-field as well. machines that compute proper pixel coverage for points and lines. Because the architecture is a direct extension of current An early example is the raster-based Evans and Sutherland PS- workstation rendering technology, we begin by discussing the 390 [E&S 87], whose design goal was to duplicate the point and performance and quality characteristics of that technology. The line quality of its calligraphic predecessors. A simpler and less problem of spatial aliasing is then discussed, and the effective line drawing algorithm is implemented by the Silicon Accumulation Buffer is shown to be a desirable solution. Finally Graphics GT system [Akeley 88]. This solution offers a great the generality of the Accumulation Buffer is explored, improvement over aliased lines, but still displays slope and concentrating on its application to the problems of motion blur, endpoint related anomalies. Both these point and line solutions, depth-of-field, and soft shadows. and all others known to the authors, rely on the following two observations: CR Categories and Subject Descriptors: 1.3.1 [Computer 1. Pixel coverages are relatively easy to compute, because the Graphics]: Hardware Architecture - Raster display devices; 1.3.3 screen geometry of the scan converted primitive is regular [ComPuter Graphics]: Picture/Image Generation - display and predictable. algorithms; 1.3.7 [Computer Graphics]: Three-Dimensional Graphics and Realism - Color, shading, shadowing and texture. 2. The quality of intersections is relatively less important than the quality of background-abutting edges, both because Additional Key Words and Phrases: Accumulation buffer, background-abutting edges predominate, and because antialiasing, motion blur, depth of field, soft shadows, stochastic intersections are typically between unrelated points or lines. sampling. Neither of these assumptions is correct for polygons. Because vertexes and narrow areas are neither regular nor predictable, it is difficult to compute correct pixel coverage during polygon scan conversion. Further, polygons frequently share edges with related polygons, and intersect unrelated polygons. Still, there are some examples of workstation-class machines that attempt these calculations. The Pixel Machine [Potmesil 89] takes the brute- force approach of oversampling and convolution with an arbitrary filter. While effective, this approach does not map nicely onto conventional scan-conversion hardware (the Pixel Machine scan conversion system is an array of general purpose processors). The Permission to copy without fee all or part of this material is granted Graphicon 2000 [Star 89] implements an approximation of an A- provided that the copies are not made or distributed for direct buffer [Carpenter 84], including hardware that computes a 4x4 commercial advantage, the ACM copyright notice and the title of the coverage mask for each pixel. This implementation suffers from publicationand its date appear, and notice is given that copying is by permission of the Associationfor ComputingMachinery. To copy limited (fixed) resolution and errors at polygon vertexes and edges otherwise, or to republish, requires a fee and/or specific permission. (when the polygon is very thin). @1990 ACM-0-89791-344-2/90/008/0309 $00.75 309 O SIGGRAPH '90, Dallas, August 6-10, 1990 We sought a polygon antialiasing solution with the following 3. Sloppy iteration. Insufficient accuracy is maintained during properties: edge iteration. Slopes and initial values of parameters are • Compatibility. The solution should leverage the capabil!ties not corrected for the subpixel locations of vertexes, spans or of a contemporary scan conversion system. It should be scans. orthogonal to the features already present, including surface Early Silicon Graphics machines [SGI 85] and Hewlett Packard generation, blending, texture mapping, and interactive graphics systems [Swanson 86] were guilty of all thr~e errors. constructive solid geometry generation. The Silicon Graphics GT graphics system, first shipped early in • High Quality. There should be no limit to the quality of the 1988, addressed issues 2 and 3, but still forced transformed result obtained. Thus, for example, fixed size masks, and coordinates to the nearest pixel center [Akeley 88]. More recently algorithms that store finite amounts of depth or color data per shipped machines, including the Silicon Graphics Personal Iris pixel, could not be used. and the Stellar GSI000 [Apgar 88], correctly address all three concerns. The Pixel-Planes system [Fuchs 85] is an early example • Smooth performance~quality tradeoff Not only must the of an architecture that implements accurate polygon sampling. quality of the result be allowed to increase without bound, but it also must decrease smoothly toward an acceptable 2.3 Point Sampling minimum. As quality is decreased, performance must We refer to a scan conversion algorithm that rigorously selects increase toward a maximum that is competitive with other pixels for inclusion, and rigorously computes parameter values at contemporary architectures. each pixel, as a Point Sampling algorithm. Such rigor is most First we describe the performance and quality characteristics of easily defined for triangles. The requirements are: the current generation of workstation graphics systems. Then we 1. The projected vertexes of the triangle must not be perturbed describe an architecture that extends these characteristics to during the scan conversion proc,Sss. include polygon antialiasing. Finally we discuss the additional system features that result from the generality of our solution. 2. Pixels must be chosen for inclusion in the triangle scan conversion based on whether their infinitely small sample 2. Current Architectures point is inside or outside the exact triangle boundary. A fair test must be established for pixets whose sample point is The problem of correctly sampling, and thus antialiasing polygons exactly on the triangle boundary. has been solved many times in many ways. Our concern here is to solve it in a manner that complements the operation of a 3. Parameter values must be assigned to each pixel based on contemporary high-performance scan conversion system. We exact calculation at the infinitely small sample point. Such must first be familiar with the properties of such a system. exact calculation is easily defined for triangles as the solution of the plane equation specified by the parameter values at the triangle's three vertexes. 2.1 Polygon Performance The most obvious trend in high-performance workstation graphics While Point Sampling can require significantly more arithmetic is toward the capability of rendering lots of small polygons per than less rigorous sampling, it has numerous benefits. The Point second. Numbers for previous generation machines reached the Sampling pixel inclusion property insures that adjacent polygons 100,000 to 150,000 range [Akeley 88, Apgar 88]. The recently neither share nor omit any pixels along their common border. introduced Silicon Graphics 4D VGX raises this number to Thus algorithms that count on the number of times a pixel is 750,000 RGB lighted, Gouraud shaded, z-buffered, connected drawn, such as transparency and constructive solid geometry triangles per second, and to 1,000,000 per second when the [Goldfeather 86], operate correctly. Redraws of single-buffered triangles are fiat shaded. images also are much less "noisy", because pixels change color less often. Substantial hardware resources are dedicated toward achieving these impressive polygon rates. We would like to leverage this The planar sampling inherent in Point Sampling allows polygon investment when drawing antialiased polygons. intersections to be Z-buffered accurately, resulting in smooth transitions from one polygon to the other. Likewise, smooth shaded polygons show no shear artifacts, such as those illustrated 2.2 Sampling Quality at the bottom of Figure 1. A second trend is that toward improved sampling quality. Finally, as we will see in the following section, the accuracy Traditional workstation scan conversion